public override void DoWork(string[] args)
{
var folder = args[1];
var model = ModelIO.Load(folder);
DoWork(model);
ModelIO.Save(model);
}
public override void DoWork(string[] args)
{
if (args.Length < 3)
{
throw (new ArgumentException(String.Format("Insufficient args")));
}
var folder = args[1];
AssemblerMode mode;
if (!Enum.TryParse(args[2], true, out mode))
{
throw (new ArgumentException(String.Format("Unknown mode: {0}", args[2])));
}
var print = mode == AssemblerMode.Print;
var model = ModelIO.Load(folder);
DoWork(model, print);
// ModelIO.Save(model);
}
static void Main(string[] args)
{
var path = ModelIO.DebugSubdir("debug\\20");
Directory.CreateDirectory(path);
File.Delete(path + "\\montage.v3");
var model = ModelIO.Load(path);
for (int i = 0; i < 1000; i++)
{
model.SetChunkMode(i * 3000, (i % 2 == 0)?Mode.Face:Mode.Screen, false);
}
for (int i = 0; i < 1000; i++)
{
model.Montage.Intervals.Add(new Interval {
StartTime = i * 3000, EndTime = i * 3000 + 500, HasVoice = false
});
model.Montage.Intervals.Add(new Interval {
StartTime = i * 3000 + 500, EndTime = i * 3000 + 3000, HasVoice = true
});
}
ModelIO.Save(model);
}
public static void Run()
{
//訓練回数
const int learningCount = 10000;
//訓練データ
Real[][] trainData =
{
new Real[] { 0, 0 },
new Real[] { 1, 0 },
new Real[] { 0, 1 },
new Real[] { 1, 1 }
};
//訓練データラベル
Real[][] trainLabel =
{
new Real[] { 0 },
new Real[] { 1 },
new Real[] { 1 },
new Real[] { 0 }
};
//ネットワークの構成は FunctionStack に書き連ねる
FunctionStack nn = new FunctionStack(
new Linear(2, 2, name: "l1 Linear"),
new Sigmoid(name: "l1 Sigmoid"),
new Linear(2, 2, name: "l2 Linear")
);
//optimizerを宣言
nn.SetOptimizer(new MomentumSGD());
//訓練ループ
Console.WriteLine("Training...");
for (int i = 0; i < learningCount; i++)
{
for (int j = 0; j < trainData.Length; j++)
{
//訓練実行時にロス関数を記述
Trainer.Train(nn, trainData[j], trainLabel[j], new SoftmaxCrossEntropy());
}
}
//訓練結果を表示
Console.WriteLine("Test Start...");
foreach (Real[] input in trainData)
{
NdArray result = nn.Predict(input)[0];
int resultIndex = Array.IndexOf(result.Data, result.Data.Max());
Console.WriteLine(input[0] + " xor " + input[1] + " = " + resultIndex + " " + result);
}
//学習の終わったネットワークを保存
ModelIO.Save(nn, "test.nn");
//学習の終わったネットワークを読み込み
Function testnn = ModelIO.Load("test.nn");
Console.WriteLine("Test Start...");
foreach (Real[] input in trainData)
{
NdArray result = testnn.Predict(input)[0];
int resultIndex = Array.IndexOf(result.Data, result.Data.Max());
Console.WriteLine(input[0] + " xor " + input[1] + " = " + resultIndex + " " + result);
}
}
public static void Run()
{
RILogManager.Default?.SendDebug("MNIST Data Loading...");
MnistData mnistData = new MnistData(28);
RILogManager.Default?.SendDebug("Training Start...");
int neuronCount = 28;
FunctionStack nn = new FunctionStack("Test19",
new Linear(true, neuronCount * neuronCount, N, name: "l1 Linear"), // L1
new BatchNormalization(true, N, name: "l1 BatchNorm"),
new LeakyReLU(slope: 0.000001, name: "l1 LeakyReLU"),
new Linear(true, N, N, name: "l2 Linear"), // L2
new BatchNormalization(true, N, name: "l2 BatchNorm"),
new LeakyReLU(slope: 0.000001, name: "l2 LeakyReLU"),
new Linear(true, N, N, name: "l3 Linear"), // L3
new BatchNormalization(true, N, name: "l3 BatchNorm"),
new LeakyReLU(slope: 0.000001, name: "l3 LeakyReLU"),
new Linear(true, N, N, name: "l4 Linear"), // L4
new BatchNormalization(true, N, name: "l4 BatchNorm"),
new LeakyReLU(slope: 0.000001, name: "l4 LeakyReLU"),
new Linear(true, N, N, name: "l5 Linear"), // L5
new BatchNormalization(true, N, name: "l5 BatchNorm"),
new LeakyReLU(slope: 0.000001, name: "l5 LeakyReLU"),
new Linear(true, N, N, name: "l6 Linear"), // L6
new BatchNormalization(true, N, name: "l6 BatchNorm"),
new LeakyReLU(slope: 0.000001, name: "l6 LeakyReLU"),
new Linear(true, N, N, name: "l7 Linear"), // L7
new BatchNormalization(true, N, name: "l7 BatchNorm"),
new LeakyReLU(slope: 0.000001, name: "l7 ReLU"),
new Linear(true, N, N, name: "l8 Linear"), // L8
new BatchNormalization(true, N, name: "l8 BatchNorm"),
new LeakyReLU(slope: 0.000001, name: "l8 LeakyReLU"),
new Linear(true, N, N, name: "l9 Linear"), // L9
new BatchNormalization(true, N, name: "l9 BatchNorm"),
new PolynomialApproximantSteep(slope: 0.000001, name: "l9 PolynomialApproximantSteep"),
new Linear(true, N, N, name: "l10 Linear"), // L10
new BatchNormalization(true, N, name: "l10 BatchNorm"),
new PolynomialApproximantSteep(slope: 0.000001, name: "l10 PolynomialApproximantSteep"),
new Linear(true, N, N, name: "l11 Linear"), // L11
new BatchNormalization(true, N, name: "l11 BatchNorm"),
new PolynomialApproximantSteep(slope: 0.000001, name: "l11 PolynomialApproximantSteep"),
new Linear(true, N, N, name: "l12 Linear"), // L12
new BatchNormalization(true, N, name: "l12 BatchNorm"),
new PolynomialApproximantSteep(slope: 0.000001, name: "l12 PolynomialApproximantSteep"),
new Linear(true, N, N, name: "l13 Linear"), // L13
new BatchNormalization(true, N, name: "l13 BatchNorm"),
new PolynomialApproximantSteep(slope: 0.000001, name: "l13 PolynomialApproximantSteep"),
new Linear(true, N, N, name: "l14 Linear"), // L14
new BatchNormalization(true, N, name: "l14 BatchNorm"),
new PolynomialApproximantSteep(slope: 0.000001, name: "l14 PolynomialApproximantSteep"),
new Linear(true, N, 10, name: "l15 Linear") // L15
);
nn.SetOptimizer(new AdaGrad());
//nn.SetOptimizer(new Adam());
RunningStatistics stats = new RunningStatistics();
Histogram lossHistogram = new Histogram();
Histogram accuracyHistogram = new Histogram();
Real totalLoss = 0;
long totalLossCounter = 0;
Real highestAccuracy = 0;
Real bestLocalLoss = 0;
Real bestTotalLoss = 0;
// First skeleton save
ModelIO.Save(nn, nn.Name);
for (int epoch = 0; epoch < 1; epoch++)
{
RILogManager.Default?.SendDebug("epoch " + (epoch + 1));
RILogManager.Default?.ViewerSendWatch("epoch", (epoch + 1));
for (int i = 1; i < TRAIN_DATA_COUNT + 1; i++)
{
RILogManager.Default?.SendInformation("batch count " + i + "/" + TRAIN_DATA_COUNT);
TestDataSet datasetX = mnistData.GetRandomXSet(BATCH_DATA_COUNT, 28, 28);
Real sumLoss = Trainer.Train(nn, datasetX.Data, datasetX.Label, new SoftmaxCrossEntropy());
totalLoss += sumLoss;
totalLossCounter++;
stats.Push(sumLoss);
lossHistogram.AddBucket(new Bucket(-10, 10));
accuracyHistogram.AddBucket(new Bucket(-10.0, 10));
if (sumLoss < bestLocalLoss && sumLoss != Double.NaN)
{
bestLocalLoss = sumLoss;
}
if (stats.Mean < bestTotalLoss && sumLoss != Double.NaN)
{
bestTotalLoss = stats.Mean;
}
try
{
lossHistogram.AddData(sumLoss);
}
catch (Exception)
{
}
if (i % 20 == 0)
{
RILogManager.Default?.SendDebug("\nbatch count " + i + "/" + TRAIN_DATA_COUNT);
RILogManager.Default?.SendDebug("Total/Mean loss " + stats.Mean);
RILogManager.Default?.SendDebug("local loss " + sumLoss);
RILogManager.Default?.SendInformation("batch count " + i + "/" + TRAIN_DATA_COUNT);
RILogManager.Default?.ViewerSendWatch("batch count", i);
RILogManager.Default?.ViewerSendWatch("Total/Mean loss", stats.Mean);
RILogManager.Default?.ViewerSendWatch("local loss", sumLoss);
RILogManager.Default?.SendDebug("");
RILogManager.Default?.SendDebug("Testing...");
TestDataSet datasetY = mnistData.GetRandomYSet(TEST_DATA_COUNT, 28);
Real accuracy = Trainer.Accuracy(nn, datasetY.Data, datasetY.Label);
if (accuracy > highestAccuracy)
{
highestAccuracy = accuracy;
}
RILogManager.Default?.SendDebug("Accuracy: " + accuracy);
RILogManager.Default?.ViewerSendWatch("Accuracy", accuracy);
try
{
accuracyHistogram.AddData(accuracy);
}
catch (Exception)
{
}
}
}
}
RILogManager.Default?.SendDebug("Best Accuracy: " + highestAccuracy);
RILogManager.Default?.SendDebug("Best Total Loss " + bestTotalLoss);
RILogManager.Default?.SendDebug("Best Local Loss " + bestLocalLoss);
RILogManager.Default?.ViewerSendWatch("Best Accuracy:", highestAccuracy);
RILogManager.Default?.ViewerSendWatch("Best Total Loss", bestTotalLoss);
RILogManager.Default?.ViewerSendWatch("Best Local Loss", bestLocalLoss);
// Save all with training data
ModelIO.Save(nn, nn.Name);
}
public static void Run()
{
const int learningCount = 10000;
Real[][] trainData =
{
new Real[] { 0, 0 },
new Real[] { 1, 0 },
new Real[] { 0, 1 },
new Real[] { 1, 1 }
};
//Training data label
Real[][] trainLabel =
{
new Real[] { 0 },
new Real[] { 1 },
new Real[] { 1 },
new Real[] { 0 }
};
//Network configuration is written in FunctionStack
FunctionStack nn = new FunctionStack(
new Linear(2, 2, name: "l1 Linear"),
new Sigmoid(name: "l1 Sigmoid"),
new Linear(2, 2, name: "l2 Linear")
);
//optimizer
nn.SetOptimizer(new MomentumSGD());
Console.WriteLine("Training...");
for (int i = 0; i < learningCount; i++)
{
for (int j = 0; j < trainData.Length; j++)
{
//Describe the loss function at training execution
Trainer.Train(nn, trainData[j], trainLabel[j], new SoftmaxCrossEntropy());
}
}
//Show training results
Console.WriteLine("Test Start...");
foreach (Real[] input in trainData)
{
NdArray result = nn.Predict(input)[0];
int resultIndex = Array.IndexOf(result.Data, result.Data.Max());
Console.WriteLine(input[0] + " xor " + input[1] + " = " + resultIndex + " " + result);
}
//Save network after learning
ModelIO.Save(nn, "test.nn");
//Load the network after learning
FunctionStack testnn = ModelIO.Load("test.nn");
Console.WriteLine("Test Start...");
foreach (Real[] input in trainData)
{
NdArray result = testnn.Predict(input)[0];
int resultIndex = Array.IndexOf(result.Data, result.Data.Max());
Console.WriteLine(input[0] + " xor " + input[1] + " = " + resultIndex + " " + result);
}
}
const int N = 30; //It operates at 1000 similar to the reference link but it is slow at the CPU
public static void Run()
{
RILogManager.Default?.SendDebug("MNIST Data Loading...");
MnistData mnistData = new MnistData(28);
RILogManager.Default?.SendDebug("Training Start...");
//Writing the network configuration in FunctionStack
FunctionStack nn = new FunctionStack("Test7",
new Linear(true, 28 * 28, N, name: "l1 Linear"), // L1
new BatchNormalization(true, N, name: "l1 BatchNorm"),
new ReLU(name: "l1 ReLU"),
new Linear(true, N, N, name: "l2 Linear"), // L2
new BatchNormalization(true, N, name: "l2 BatchNorm"),
new ReLU(name: "l2 ReLU"),
new Linear(true, N, N, name: "l3 Linear"), // L3
new BatchNormalization(true, N, name: "l3 BatchNorm"),
new ReLU(name: "l3 ReLU"),
new Linear(true, N, N, name: "l4 Linear"), // L4
new BatchNormalization(true, N, name: "l4 BatchNorm"),
new ReLU(name: "l4 ReLU"),
new Linear(true, N, N, name: "l5 Linear"), // L5
new BatchNormalization(true, N, name: "l5 BatchNorm"),
new ReLU(name: "l5 ReLU"),
new Linear(true, N, N, name: "l6 Linear"), // L6
new BatchNormalization(true, N, name: "l6 BatchNorm"),
new ReLU(name: "l6 ReLU"),
new Linear(true, N, N, name: "l7 Linear"), // L7
new BatchNormalization(true, N, name: "l7 BatchNorm"),
new ReLU(name: "l7 ReLU"),
new Linear(true, N, N, name: "l8 Linear"), // L8
new BatchNormalization(true, N, name: "l8 BatchNorm"),
new ReLU(name: "l8 ReLU"),
new Linear(true, N, N, name: "l9 Linear"), // L9
new BatchNormalization(true, N, name: "l9 BatchNorm"),
new ReLU(name: "l9 ReLU"),
new Linear(true, N, N, name: "l10 Linear"), // L10
new BatchNormalization(true, N, name: "l10 BatchNorm"),
new ReLU(name: "l10 ReLU"),
new Linear(true, N, N, name: "l11 Linear"), // L11
new BatchNormalization(true, N, name: "l11 BatchNorm"),
new ReLU(name: "l11 ReLU"),
new Linear(true, N, N, name: "l12 Linear"), // L12
new BatchNormalization(true, N, name: "l12 BatchNorm"),
new ReLU(name: "l12 ReLU"),
new Linear(true, N, N, name: "l13 Linear"), // L13
new BatchNormalization(true, N, name: "l13 BatchNorm"),
new ReLU(name: "l13 ReLU"),
new Linear(true, N, N, name: "l14 Linear"), // L14
new BatchNormalization(true, N, name: "l14 BatchNorm"),
new ReLU(name: "l14 ReLU"),
new Linear(true, N, 10, name: "l15 Linear") // L15
);
nn.SetOptimizer(new AdaGrad());
for (int epoch = 0; epoch < 3; epoch++)
{
Real totalLoss = 0;
long totalLossCounter = 0;
//Run the batch
for (int i = 1; i < TRAIN_DATA_COUNT + 1; i++)
{
RILogManager.Default?.SendDebug("epoch " + (epoch + 1) + " of 3, Batch " + i + " of " + TRAIN_DATA_COUNT);
//Get data randomly from training data
TestDataSet datasetX = mnistData.GetRandomXSet(BATCH_DATA_COUNT, 28, 28);
//Learn
Real sumLoss = Trainer.Train(nn, datasetX.Data, datasetX.Label, new SoftmaxCrossEntropy());
totalLoss += sumLoss;
totalLossCounter++;
if (i % 20 == 0)
{
RILogManager.Default?.SendDebug("batch count " + i + "/" + TRAIN_DATA_COUNT);
RILogManager.Default?.SendDebug("total loss " + totalLoss / totalLossCounter);
RILogManager.Default?.SendDebug("local loss " + sumLoss);
RILogManager.Default?.SendDebug("Testing random data...");
//Get data randomly from test data
TestDataSet datasetY = mnistData.GetRandomYSet(TEST_DATA_COUNT, 28);
//Run the test
Real accuracy = Trainer.Accuracy(nn, datasetY.Data, datasetY.Label);
RILogManager.Default?.SendDebug("Test Accuracy: " + accuracy);
}
}
}
ModelIO.Save(nn, "Test7.nn");
RILogManager.Default?.SendDebug(nn.Describe());
}
public static void Run()
{
int neuronCount = 28;
RILogManager.Default?.SendDebug("MNIST Data Loading...");
MnistData mnistData = new MnistData(neuronCount);
RILogManager.Default.SendInformation("Training Start, creating function stack.");
SortedFunctionStack nn = new SortedFunctionStack();
SortedList <Function> functions = new SortedList <Function>();
ParallelOptions po = new ParallelOptions();
po.MaxDegreeOfParallelism = 4;
for (int x = 0; x < numLayers; x++)
{
Application.DoEvents();
functions.Add(new Linear(true, neuronCount * neuronCount, N, name: $"l{x} Linear"));
functions.Add(new BatchNormalization(true, N, name: $"l{x} BatchNorm"));
functions.Add(new ReLU(name: $"l{x} ReLU"));
RILogManager.Default.ViewerSendWatch("Total Layers", (x + 1));
}
;
RILogManager.Default.SendInformation("Adding Output Layer");
Application.DoEvents();
nn.Add(new Linear(true, N, 10, noBias: false, name: $"l{numLayers + 1} Linear"));
RILogManager.Default.ViewerSendWatch("Total Layers", numLayers);
RILogManager.Default.SendInformation("Setting Optimizer to AdaGrad");
nn.SetOptimizer(new AdaGrad());
Application.DoEvents();
RunningStatistics stats = new RunningStatistics();
Histogram lossHistogram = new Histogram();
Histogram accuracyHistogram = new Histogram();
Real totalLoss = 0;
long totalLossCounter = 0;
Real highestAccuracy = 0;
Real bestLocalLoss = 0;
Real bestTotalLoss = 0;
for (int epoch = 0; epoch < 3; epoch++)
{
RILogManager.Default?.SendDebug("epoch " + (epoch + 1));
RILogManager.Default.SendInformation("epoch " + (epoch + 1));
RILogManager.Default.ViewerSendWatch("epoch", (epoch + 1));
Application.DoEvents();
for (int i = 1; i < TRAIN_DATA_COUNT + 1; i++)
{
Application.DoEvents();
TestDataSet datasetX = mnistData.GetRandomXSet(BATCH_DATA_COUNT, neuronCount, neuronCount);
Real sumLoss = Trainer.Train(nn, datasetX.Data, datasetX.Label, new SoftmaxCrossEntropy());
totalLoss += sumLoss;
totalLossCounter++;
stats.Push(sumLoss);
lossHistogram.AddBucket(new Bucket(-10, 10));
accuracyHistogram.AddBucket(new Bucket(-10.0, 10));
if (sumLoss < bestLocalLoss && !double.IsNaN(sumLoss))
{
bestLocalLoss = sumLoss;
}
if (stats.Mean < bestTotalLoss && !double.IsNaN(sumLoss))
{
bestTotalLoss = stats.Mean;
}
try
{
lossHistogram.AddData(sumLoss);
}
catch (Exception)
{
}
if (i % 20 == 0)
{
RILogManager.Default.ViewerSendWatch("Batch Count ", i);
RILogManager.Default.ViewerSendWatch("Total/Mean loss", stats.Mean);
RILogManager.Default.ViewerSendWatch("Local loss", sumLoss);
RILogManager.Default.SendInformation("Batch Count " + i + "/" + TRAIN_DATA_COUNT + ", epoch " + epoch + 1);
RILogManager.Default.SendInformation("Total/Mean loss " + stats.Mean);
RILogManager.Default.SendInformation("Local loss " + sumLoss);
Application.DoEvents();
RILogManager.Default?.SendDebug("Testing...");
TestDataSet datasetY = mnistData.GetRandomYSet(TEST_DATA_COUNT, 28);
Real accuracy = Trainer.Accuracy(nn, datasetY?.Data, datasetY.Label);
if (accuracy > highestAccuracy)
{
highestAccuracy = accuracy;
}
RILogManager.Default?.SendDebug("Accuracy: " + accuracy);
RILogManager.Default.ViewerSendWatch("Best Accuracy: ", highestAccuracy);
RILogManager.Default.ViewerSendWatch("Best Total Loss ", bestTotalLoss);
RILogManager.Default.ViewerSendWatch("Best Local Loss ", bestLocalLoss);
Application.DoEvents();
try
{
accuracyHistogram.AddData(accuracy);
}
catch (Exception)
{
}
}
}
}
ModelIO.Save(nn, Application.StartupPath + "\\test20.nn");
RILogManager.Default?.SendDebug("Best Accuracy: " + highestAccuracy);
RILogManager.Default?.SendDebug("Best Total Loss " + bestTotalLoss);
RILogManager.Default?.SendDebug("Best Local Loss " + bestLocalLoss);
RILogManager.Default.ViewerSendWatch("Best Accuracy: ", highestAccuracy);
RILogManager.Default.ViewerSendWatch("Best Total Loss ", bestTotalLoss);
RILogManager.Default.ViewerSendWatch("Best Local Loss ", bestLocalLoss);
}
public static void Run()
{
const int learningCount = 10000;
Real[][] trainData =
{
new Real[] { 0, 0 },
new Real[] { 1, 0 },
new Real[] { 0, 1 },
new Real[] { 1, 1 }
};
Real[][] trainLabel =
{
new Real[] { 0 },
new Real[] { 1 },
new Real[] { 1 },
new Real[] { 0 }
};
bool verbose = true;
FunctionStack nn = new FunctionStack("Test1",
new Linear(verbose, 2, 2, name: "l1 Linear"),
new Sigmoid(name: "l1 Sigmoid"),
new Linear(verbose, 2, 2, name: "l2 Linear"));
nn.SetOptimizer(new MomentumSGD());
Info("Training...");
for (int i = 0; i < learningCount; i++)
{
for (int j = 0; j < trainData.Length; j++)
{
Trainer.Train(nn, trainData[j], trainLabel[j], new SoftmaxCrossEntropy());
}
}
Info("Test Start...");
foreach (Real[] input in trainData)
{
NdArray result = nn.Predict(true, input)?[0];
int resultIndex = Array.IndexOf(result?.Data, result.Data.Max());
Info($"{input[0]} xor {input[1]} = {resultIndex} {result}");
}
Info("Saving Model...");
ModelIO.Save(nn, "test.nn");
Info("Loading Model...");
FunctionStack testnn = ModelIO.Load("test.nn");
Info(testnn.Describe());
Info("Test Start...");
foreach (Real[] input in trainData)
{
NdArray result = testnn?.Predict(true, input)?[0];
int resultIndex = Array.IndexOf(result?.Data, result?.Data.Max());
Info($"{input[0]} xor {input[1]} = {resultIndex} {result}");
}
}
